A year ago an Nvidia Grace Blackwell went out on long-term contracts at $5 per GPU hour. Jensen Huang says the spot price today is $16.
The week's argument has been about whether the leading artificial-intelligence labs should slow down. Huang answered by pricing the machines instead of debating the risk, and on his numbers the rental market has moved faster than the buildout.
"So yeah, the return on invested capital is about one year right now."
Huang runs the company whose chips every frontier lab trains on, and was at Salesforce's Dreamforce conference the day after taking a phone call from the President of the United States in the middle of a podcast taping.
The full segment is covered here so you can skip it.
Here are the 8 takeaways that matter.
👤 Guest: Jensen Huang, Founder, President and CEO of Nvidia
🎙️ Host: Jim Cramer, host of Mad Money and manager of the CNBC Investing Club's charitable trust
🧩 Other segments: Marc Benioff of Salesforce and Sarah Friar of OpenAI
📰 Published: 15 September 2026 on the Mad Money podcast (CNBC)
🟣 Apple Podcasts | ⏱️ 13 min
Key Takeaways
A gigawatt of Nvidia AI factories costs about $50 to $60 billion and rents for about $50 billion a year
That is a return on invested capital of roughly one year, on his own arithmetic
The rental price of the same hardware has more than tripled inside a year
$16 per GPU hour on the spot market against $5 on contracts signed a year ago
Safety is an engineering problem, and no new law is needed to make companies do it
His test is the release gate: keep testing and keep engineering until the product is ready
Open models went from 30% of tokens generated to almost 70% in under a year
Total token generation demand is up 25 times over the same stretch
Safety testing becomes a third class of large system, alongside training and inference
He counts it as new demand for data centers rather than a brake on them
Nvidia compute is the first computing system that can be pledged as collateral
Because it runs every model and every stage of the lifecycle, it stays fungible and durable
Nobody can outsource intelligence, which is why every country and company still builds its own
Salesforce is building proprietary agents on Nvidia's Nemotron models
He calls the story that productivity destroys jobs "just wrong"
His mechanism is ambition: a more productive company attempts more, and needs more people to do it
1. Safety Is Engineering
Cramer opened the segment at Dreamforce by saying he had been to his trust-and-estate lawyer to ask what he could do for his children if he knew he was going to die in 2030, and that the lawyer questioned the premise.
Huang rejected it flatly. "We're not going to die in 2030." His reason is a head count of participants: thousands of companies are building AI rather than two, and between them they will build guardrails and invent safety and security technology
He says the three goals are compatible rather than traded off against each other. AI will be incredibly useful, incredibly helpful and incredibly safe — "We could have all of that at the same time."
The formulation he repeated is that this is a build problem, not a policy one. "Safety is an engineering problem. Testing is an engineering problem."
His prescription is a release gate. "We should create products and properly test them. And if they're not ready to be released, just hold on to it and keep testing it and keep engineering." Cramer's reaction was that it is common sense and that he is a believer
Cramer had framed the week for viewers before the interview started: a tug of war over AI in which Anthropic and OpenAI, which he called the two private kingpins, want to slow development while the rest of the food chain wants to keep building at a fast clip
Cramer's own position was that the money gets spent either way. He said plenty will be spent on technology and there is no slowdown there, despite what many people on Wall Street think
2. A One-Year Payback
Cramer reminded Huang that two years ago he had said buyers would make four times what they expected out of Nvidia equipment, and that customers are making money on it now but do not like to talk about it.
Huang dated the change to the last six months, when on his account AI became super useful. Because it is useful, he said, people are building compute systems — Nvidia AI factories — like crazy
The build cost and the rent, in his own figures: "Each gigawatt of Nvidia AI factories is about $50 to $60 billion." The rentals of that same one-gigawatt factory run about $50 billion a year
That ratio is the claim the rest of the segment rests on. "So yeah, the return on invested capital is about one year right now."
The second half of the return is how long the asset lasts. Because Nvidia's fungibility and durability are so great, he said, buyers get to use the equipment for many years, well beyond five and six
His one-line explanation for the buildout is that the arithmetic works: "That's why everybody's building."
3. Build Your Own Intelligence
Cramer put it to him that the closed frontier models can apparently make three times that, and asked whether both open and closed models are needed.
Huang's answer separates renting from owning. Rent and use the best models available as a product or a service, he said — and then every single country, every single company and every single industry must also build its own intelligence
The question he used to close the point: "How can you outsource intelligence to somebody else completely?"
The Salesforce deal is his worked example. Salesforce will use Nvidia's Nemotron models and its agent technology to build proprietary agents that sit on top of Salesforce
The second half of that partnership is the data layer, and it reaches past Salesforce. He said the data-processing layer sits in Salesforce, in Databricks and in Snowflake, that Nvidia accelerates data processing across all of it, and that the result is data processing at lightning speeds with agents running on top
His conclusion about the company hosting the conference: "And so this way Salesforce is an AI company as well."
4. The Loom Story Is Wrong
Cramer offered the standard industrial-revolution comparison — the loom put a lot of seamstresses out of business, it was a tough game for them, and four years later there were ten times the jobs.
Huang refused the comparison rather than accepting it. "You know that storytelling is unfortunate, unnecessary and quite frankly, just wrong."
His mechanism is ambition rather than head count. When a company becomes more productive and is powered by artificial intelligence, he said, it becomes more ambitious, and that ambition lets it imagine and attempt greater things
The evidence he cited was the company hosting him. Salesforce was a data layer and an application layer and is now an agent layer on top of that; he said Marc Benioff is more ambitious about the company's future than at any time he has seen, and that AI is what raised the ambition
He extended the claim past Salesforce to Salesforce's own customers and to every single industry
On software engineering he made a generational argument. He called it completely illogical that engineering would stop being necessary because coding agents exist, said he grew up in a generation of engineers who had no software and built prototypes with other tools, and finished: "There will be engineering after AI."
5. The Call From the President
Cramer asked about the podcast appearance in which Huang took a phone call in the middle of the interview, and whether it had been planned. It had not.
Huang's account is that he was on stage and saw his team waving a phone in the air out of the corner of his eye. As they came closer they said it was the President
The rule he applied is short. "Take it when the president calls, you pick it up."
What the call was about was the President's own posts. Huang said the President wanted to talk about some of the tweets he had put out; when Huang mentioned he was on stage with the hosts, the President asked to be put on so he could talk to everybody
Cramer's follow-up was about the word the President used on the call. It was hoax, and that is the part a lot of people picked up — but Cramer said the panel's own consensus was that some things are made up, and asked Huang how you stop that
6. No New Laws Needed
His first answer to how you stop it was to concede the problem. "AI safety is a real thing."
The obligation he accepts sits on release, not on speed. Engineering products so they are safe for the world to use is a real thing, and "And companies should innovate as fast as possible, but they should never innovate so fast as to release unsafe products."
What he rejects is new statute. The idea that new laws, new antitrust laws or new regulations are needed so that these companies can do their fundamental engineering properly before releasing products is, in his words, "just completely unnecessary"
He read that as the President's position as well — that we do not need new laws
His reason is that the existing ones already cover it. "We got plenty of regulations that govern the reliability and the functionality of products."
7. Testing Is a Third System
Cramer described the previous day as weird: the stock opened down five and a half percent on the view that OpenAI will not come public and that Nvidia would therefore miss its quarter. He put the stock at 14 times fiscal 2028 earnings, a year that starts in January, and called it the best investment in the world.
Huang answered on coverage rather than on valuation. Nvidia is the only platform that runs every closed model, and is gaining share in closed models; it is also the only platform that runs every open model, including Nvidia's own Nemotron
Two demand figures he gave for the past year: token generation demand is up 25 times in less than a year, and the share of tokens generated by open models has gone from 30% to almost 70%
His reading of the week's news was that it changes nothing. "Nothing, nothing that happened this week is going to slow that down."
The new claim inside that answer is a third category of large system. Until now there have been training systems and inference systems; he says environmental testing labs become the third, and that they will be extremely large, because testing products for function, capability, safety and security requires massive data centers
He therefore counted the safety debate as demand rather than drag. "So we just created yet another new demand for AI."
8. Compute as Collateral
Cramer asked whether there is enough compute, and when the financial industry will start trading it.
Huang called this the big breakthrough of the last few years. "People realize now that Nvidia compute is the only system that is fungible and durable."
The reason he gives for the fungibility is coverage of the whole lifecycle. Nvidia runs every closed and open model and every stage from data processing to training to post-training to inference and now testing, and every single company uses it
That is what turns the machines into a financial asset. "And so you could use the computers as collateral." He said the evidence for it is accumulating, and called it a huge unlock for companies that cannot otherwise fund a buildout of this scale, because the asset itself becomes the security
The spot market is what moved. Many early renters signed long-term leases and contracts, and Huang said the cloud providers now regret it: a Grace Blackwell rents for $16 per GPU hour today, and "Now, some of them were rented away at $5 just a year ago."
What happens as those contracts roll off is the second-order effect. Everybody is raising prices and raising revenue forecasts, the return on invested capital is coming through, and they are buying more gear and putting up more infrastructure
His summary of the loop: "And so I think this flywheel is really, really flying."
Bonus Insights
Cramer's market read at the top of the show: oil prices surged, bond yields took off again and stocks got hammered, with the Dow tumbling 328 points, the S&P 500 off 0.45% and the Nasdaq losing 0.78%
Why the show was in San Francisco at all. Cramer said he has been coming to Dreamforce for more than a decade because for a few days a year it becomes the center of the universe, bringing together the people building the future and the companies spending billions to make it happen
Huang's own description of the venue was commercial. He called it the world's epicenter of selling, and of creating opportunities for the world
The sign-off was Cramer's standing instruction on the stock, and Huang took it. Cramer said "Nvidia own it. Don't trade it," and Huang replied: "Own it don't trade it. That's what I do."
Huang's bottom line is that the AI argument should be settled by engineering and arithmetic rather than by legislation: test products properly before releasing them, and keep buying the machines, because a gigawatt of them pays for itself in about a year and can now be pledged as collateral.
Products, Companies & Tools Mentioned
Nvidia (His own company — the only platform, he says, that runs every closed and every open model and every stage of the AI lifecycle, which is what makes its compute fungible, durable and now financeable)
Nvidia Nemotron (Nvidia's own open models, which Salesforce will use with Nvidia's agent technology to build proprietary agents)
Nvidia Grace Blackwell (The system whose spot rental he puts at $16 per GPU hour against $5 on contracts signed a year ago)
Salesforce (The conference host and the partnership he used as his example of a company becoming an AI company on top of its existing data and application layers)
Databricks and Snowflake (Named with Salesforce as the data-processing layers Nvidia accelerates, so that agents run against data processed at speed)
OpenAI and Anthropic (Cramer's "two private kingpins" who want to slow development of the most powerful models, against a food chain that wants to keep building)
Watch the full episode:
If this was worth your time, send it to someone who follows the name.
Get the latest market chatter as it happens: